Writing for Humans and AI at the Same Time

How big this channel actually is

Start with the number that puts the rest in proportion. According to a large publisher traffic analysis, since July 2024 one of the biggest AI assistants has accounted for 0.02 percent of total referral traffic, and another major provider for 0.002 percent. The same analysis puts it plainly: Google sends five hundred times as many referrals from search alone, and thirteen hundred times as many once its discovery surface is included.

Meanwhile that five-hundred-times base is shrinking. The same report measured organic search traffic to news sites falling 33 percent globally between November 2024 and November 2025, and 38 percent in the United States. A separate panel-based measurement found that in the first four months of 2026, 68 percent of Google searches ended without a click, against 60.45 percent in 2024.

So the situation is not that a new channel is taking the old one's place. What is happening is that the old channel is narrowing, and the new one is orders of magnitude smaller than what the narrowing removes. One sober conclusion follows: in 2026 an AI answer engine is a visibility and credibility surface, not a traffic channel. Anyone promising traffic growth from it has either not looked at the numbers or is selling something else.

What the research says works

The literature on "generative engine optimisation" is short and uncomfortably consistent. The most rigorous peer-reviewed study tested the established tactics across 1,900 queries and concluded that most methods are not only largely ineffective but frequently harm the document's ranking. The same work adds that "traditional SEO strategies, those aiming to improve the ranking of the source, are significantly more effective".

A survey paper published in the summer of 2026 says the same from another angle: the only reproducible factor is the topical relevance of query to document, followed by where your text sits in the model's context. That same survey notes that of 54 method-and-domain combinations, exactly three produced significantly positive results, and that gains decline as more people adopt the same tactic.

What measurably helps is narrow and prosaic. The original 2023 study found that naming sources and adding verbatim quotations each raised visibility by roughly 27 percent, while keyword stuffing cost 8.3 percent. But here is the caveat that matters: those numbers apply to documents already in the model's context. They do not affect whether you get found, only whether you get quoted once you have been.

What actually matters: getting into the context at all

If the research says one thing, it is that retrieval beats phrasing. Five items belong here, and none of them is news.

• Conventional ranking. What ranks well in search is more likely to be pulled into an answer. This is the dullest and best-supported finding in the whole subject.
• Coverage of the sub-questions. Answer engines split one question into several sub-queries and search for each separately. By one measurement, text that surfaces across several of those sub-queries is 161 percent more likely to end up cited.
• Being mentioned elsewhere. The same measurement found mentions on other sites to be the second strongest correlate. What you do not write about yourself counts for more than what you do.
• Being indexable. Google's own documentation says appearing on AI surfaces has one condition: the page must be indexed and eligible to be shown with a snippet. Nothing else.
• Being reachable by the answer engine at all. That is what the second-to-last section is about, and it is where the biggest mistake is available.

What is worth doing to the text itself

Five tactics the evidence supports — or at least does not contradict, and which are better for a human reader anyway. Take that distinction seriously: the second half of this list is craft judgement, not measured result.

• Name the source inside the sentence. Do not just link it: say who measured it and when. This is one of the few measured, positive interventions, and it is also the credible form for a human reader.
• Write out the concrete figure. Price, date, quantity, version. A 2026 study found a stated price and a recent timestamp among the things without which text is routinely left out of answers.
• Make each paragraph comprehensible on its own. The model sees your text in pieces, and the piece it picks goes into the answer without context. That is a reasonable expectation — but Google explicitly states there is no requirement to break content into tiny pieces, so do not overdo it.
• Do not worry about length. In a study of 1.9 million citations, the correlation between word count and being cited was effectively zero. Long text on its own brings nothing.
• A question-and-answer block helps, but it is not magic. One large search provider's own guidance mentions clear headings, tables and FAQ sections; controlled experiments, however, measured negligible effects from formatting-only changes. Do it because it is good for the reader — not because you expect citations from it.

Three things you are sold

All three are popular, all three have been measured, and none delivered what was promised.

• The llms.txt file. In May 2026, 137,000 domains were examined: 97 percent of llms.txt files received no requests at all. And AI bots never went looking for files that did not exist — "they never go looking". Google's documentation likewise says creating one "will neither harm nor help". There is one confusing contradiction: Chrome's developer tools now check for its presence — same company, two opposite messages. For now the data is unambiguous.
• Structured data for citations. A controlled study followed 1,885 pages that added structured data over six months and matched them against comparable control pages. The result: a 4.6 percent decline on one AI surface and noise-level change on the other two. Structured data is good for many things; raising citation rates is, by measurement, not one of them.
• Keyword stuffing in any new clothing. In the study cited above, this was the single tactic that measurably hurt. Answer engines are not sensitive to word repetition.

The robots.txt, where you can genuinely go wrong

This is the section that makes the rest worth reading, because here the mistake is real and reversible. The large providers use separate bots for training and for live retrieval, and blocking the two means two different things.

Blocking the training bot signals that your content should stay out of model training data. It has no effect on whether you appear in answers. Blocking the search or retrieval bot, on the other hand, means you are removed from the answers — one provider's documentation states outright that excluded sites will not be shown in its answers. In plenty of places this gets set accidentally, in a single line.

A third trap surrounds Google. The Google-Extended token governs training for and use in Gemini models, and by its own documentation it does not affect inclusion in Google Search — which means it does not remove you from AI Overviews either. Anyone who thinks they have switched off AI Overviews with it is mistaken; snippet controls do that. And one dated change: from 15 September 2026, the largest content delivery provider will block training and agent bots by default for newly onboarding domains on pages that display ads, while leaving search bots allowed. If you are migrating a site right now, check this.

How to measure it

Carefully, because the measurement itself is unstable. The survey paper found that watching four answer engines over 45 days, the overlap between cited sources ran only between 0.34 and 0.42, and that a reliable estimate needs seven to eight repetitions of the same question. A single "I checked, and now it cites me" proves nothing.

So the usable measurement is not citation rate but two other things. First, mentions of your brand name in answers about your subject, with seven or eight repetitions, monthly. Second, the quality of the traffic that does arrive — one large analysis found visitors from AI converting 31 percent better and bouncing 33 percent less, though that figure starts from a near-zero base and those visitors are further along the decision path to begin with. Look at the channel that way — few visitors, but mature ones — and your expectations of it will be realistic.

What to take away

An AI answer engine today is a visibility surface, not a traffic channel, and most of the tactics built for it are ineffective or harmful in measurements. What works is dull: be findable, cover the sub-questions, name your sources inside the sentence, and write out the concrete figure. Expect no citations from llms.txt or structured data. But go and check your robots.txt now for whether you have blocked the retrieval bot instead of the training one — it is the only mistake in this subject with an immediate, measurable price. For when each kind of search is worth using at all, see AI search or Google.